Species Overlap and Phylogenetic Relatedness Result in Community Statistical Non‐Independence (and What to Do About It)

Statistical autocorrelation between sampling units violates independence assumptions in many analyses. Here, we used simulations and empirical analyses to demonstrate how shared evolutionary history between species and species overlap among communities leads to an insidious form of autocorrelation, termed compositional autocorrelation. We simulated compositionally autocorrelated ecosystem functions across communities and assessed the type I error, statistical power and accuracy of slope estimates from naïve linear regression models and mixed models accounting for compositional autocorrelation. Mixed models maintained lower type I error, similar or higher statistical power, and more accurate slope estimates compared to linear regression. Re‐analysing an empirical dataset, we found linear regression underestimated uncertainty in species richness effects for eight of 10 ecosystem functions. As species overlap and shared evolutionary history are common features in community data, our results highlight the need to explicitly consider compositional autocorrelation in statistical analyses to ensure correct inferences.